The New Chokepoints

The AI economy is a stack.

Eight layers, bottom to top. Silicon, model weights, training data, inference compute, embeddings, agent orchestration, trust certification, user interface.

Each layer is a candidate chokepoint — a place where, if you own it, you can tax everyone else in the stack. The question for every company in AI is: which layer are we playing for?

Layer 1 is silicon. NVIDIA. 80%+ gross margins on their data-center GPUs. The entire AI industry pays them rent, whether they like it or not.

This layer is locked. The capital intensity is too high, the foundry capacity is constrained, and CUDA is a moat that took 15 years to build. Nobody is winning this layer back in the next decade.

If you weren't there a decade ago, you're not getting in now.

Layer 2 is the foundation model weights themselves. OpenAI, Anthropic, Google. Maybe Mistral and Meta on the open-weight side.

This layer is capital-locked. Training a frontier model now costs $500M+ and rising. The handful of companies who can afford it are pulling further ahead each year.

Same conclusion: closed to new entrants in the foreseeable future. If you wanted to be here, the time was 2019.

Layers 3, 4, and 5 — training data, inference compute, embeddings — are all commoditizing fast.

Data is fragmented; everyone has some, nobody has a monopoly. Inference compute lives on AWS, Azure, and GCP, with prices falling 10× per year. Vector databases are a feature in any database product now.

These layers will exist. They will be necessary. They won't be where the money is.

Now look at Layer 6. Agent orchestration — the thing that takes a user's intent ("plan my trip"), decomposes it into sub-tasks, calls the right tools, chains the results, recovers from failures.

This layer is completely up for grabs. No dominant player. Hundreds of frameworks (LangChain, AutoGPT, OpenAI's Operator, Anthropic's Claude Agent SDK, dozens more). None of them is the winner.

Whoever wins this layer captures most of what's interesting about AI in the next five years.

And Layers 7 and 8 are even more wide-open. Who certifies that an AI agent did what it said it did? Who guarantees its outputs? Who owns the surface where the user actually meets the AI — the chat window, the agent UI, the voice interface?

These layers don't yet exist as identifiable products. Five years from now, they will be three of the largest companies in the world.

Historically, the interface layer captures more value than every layer beneath it combined. Browser (Netscape → Chrome). Smartphone (iPhone). Social feed (Facebook). Every time. The pattern doesn't change just because the stack underneath is AI.

You don't need to win all eight layers. You need to win — or credibly own a position in — exactly one of them.

— From Reshuffle, Chapter 11

The strategic mistake most companies make in 2026 is treating "AI" as undifferentiated. It isn't. It's a stack with eight floors, and each floor has wildly different dynamics — different capital requirements, different defensibility, different incumbents.

Pick your layer. Or, if you can't pick a layer, become so essential to a layer that the layer can't aggregate without you. Those are the only two safe positions.

Everyone else is paying rent to all eight.

Sangeet on this in Chapter 11 ↗